Editorial illustration for Multimodal AI Model Clef, Fine-Tuned From Qwen, Comes to Ollama's Cloud
Cloudflare's Clef Model Now Available on Ollama Cloud
Multimodal AI Model Clef, Fine-Tuned From Qwen, Comes to Ollama's Cloud
Cloudflare's 27B multimodal decision model Clef landed on Ollama October 19, 2026, giving developers a way to run it locally or through Ollama's cloud instead of only through Cloudflare's own infrastructure. Clef is fine-tuned from Qwen3.8-27B and released under the Apache 2.0 license, so the weights are open for anyone to inspect, modify, or self-host.
The model's job is narrow and specific: take a state plus a schema of typed questions, and turn that into a decision. It handles text, JSON, and images, which means it can score a screenshot, a scanned receipt, or a form photo the same way it scores plain text input. Clef is also built to work with the Jev and System One APIs, so teams already running decision pipelines on those systems have a drop-in alternative to test.
What sets Clef apart from a typical chat model is how it reads input. Rather than generating a response token by token, it processes the input once and scores every possible answer to every question in that single pass. Cloudflare says that approach pays off on its own Decision Index benchmarks.
Clef, a 27B multimodal decision model from Cloudflare, is now available on Ollama. Users can run it locally or on Ollama’s cloud. Clef turns a state and a schema of typed questions into decisions.
Why this matters
Clef's pitch is narrower than most model releases we cover, and that's the point. Instead of chasing benchmark scores, Cloudflare built something that scores every answer to every question in one pass, which matters for anyone building approval flows, form validators, or receipt-processing pipelines where latency and cost per decision actually show up on a bill. Apache 2.0 licensing and a Qwen3.8-27B base mean teams can inspect, fine-tune, or self-host without vendor lock-in, and Ollama distribution lowers the barrier to trying it locally before committing cloud spend.
We'd temper the enthusiasm with a question nobody's answered yet in Cloudflare's materials: how does Clef's one-pass scoring hold up against ambiguous schemas or adversarial inputs, like a doctored screenshot or a form with contradictory fields? Multimodal decision models are easy to demo on clean examples and harder to trust in production edge cases. Developers evaluating Clef should pressure-test it against messy, real-world documents, not just the receipts and forms used in the highlight reel, before wiring it into anything that approves or denies something automatically.
Common Questions Answered
What is Clef and what specific task is it designed to perform?
Clef is a 27B multimodal decision model developed by Cloudflare that is fine-tuned from Qwen3.8-27B. Its narrow and specific job is to take a state plus a schema of typed questions and turn that into a decision, handling text and JSON inputs to score every answer in a single pass.
Where can developers run Clef and what licensing terms apply?
Clef is now available on Ollama's platform as of October 19, 2026, allowing developers to run it locally or through Ollama's cloud infrastructure instead of only through Cloudflare's own systems. The model is released under the Apache 2.0 license, meaning the weights are open for anyone to inspect, modify, or self-host without vendor lock-in.
What are the practical use cases where Clef's decision-making capability provides value?
Clef is particularly valuable for building approval flows, form validators, and receipt-processing pipelines where latency and cost per decision directly impact operational expenses. By scoring every answer to every question in one pass, it provides efficient decision-making for applications where performance and cost efficiency are critical billing considerations.
How does Clef differ from other AI model releases in terms of its approach?
Unlike most model releases that focus on chasing benchmark scores, Clef takes a narrower, more specialized approach by being purpose-built for specific decision-making tasks. This focused design philosophy prioritizes practical efficiency and cost-effectiveness for real-world applications over broad general-purpose capabilities.
Further Reading
- Clef: A 27B Multimodal Decision Model - Hugging Face
- Clef 27B Multimodal Model - Ollama
- Clef Cloud Model - Ollama
- Clef Decision Models on the Cloudflare Blog - Cloudflare Blog